Ingest Brain

Use the brain CLI to add content to the user's hybrid-search corpus. Pick the right ingestion path based on where the content lives. Never invent content — ingest only what the user has actually pointed you at or what an MCP returned.

For searching/synthesis, see consult-brain. For authoring brand-new vault notes, see brain-authoring.

Choosing the right command

Each source with a recipe has a reference file in sources/. Before ingesting from that source, read its file in full. For Jira and GitHub the recipe is the scope, so never run the command from its table row alone.

Files and directories

brain ingest ~/Downloads/2026-Q1-OKRs.pdf --tag okrs
brain ingest-dir ~/career-docs --tag career --dry-run    # preview first
brain ingest-dir ~/career-docs --tag career --ext md,pdf

Verbs in output:

• ingested: — new row inserted
• updated: — existing row replaced in-place (content or tags changed, or --force)
• skipped: — content unchanged

--force on brain ingest re-applies even when content is byte-identical. Useful after a chunker / embedder upgrade.

brain ingest-dir has no --force flag — it always UPDATEs in-place when the file changed and skips otherwise.

To mirror deletions from a directory imported earlier, add --sync: it lists the documents whose files are gone (would delete: …, then `N of M documents under

The context cost of the MCP→stdin chain

Read this before running any MCP→stdin recipe in sources/: Krisp, Slack, Jira, and GitHub through the connector.

Ingest is the most expensive thing this skill does to your context, and it is expensive in a way the other brain skills are not. In consult-brain you choose how much text to pull in (--brief, brain recall --budget N). Here you have no such control: the MCP returns the full transcript or thread, that text lands in your context, and you echo it straight back out to brain ingest-stdin. Every ingested document transits your context twice — once inbound from the MCP, once outbound in the echo.

Measured on this corpus, a Krisp transcript is a mean of ~7,900 tokens (median 7,400; p90 15,300; largest 33,600 — n=261; reproducible from docs/audits/2026-08-11-wave2-routing-counterfactual.md). Doubled by the echo, that is ~16k tokens of context per meeting. Ten meetings in one pass is ~160k — a context window. Consequences to plan around:

• Batch small. "Ingest last week's Krisp calls" is 5–15 meetings. Do them a few at a time and report counts as you go, rather than fetching all of them before the first ingest-stdin.
• Never re-read what you just wrote. After ingest-stdin succeeds, the content is in the brain. Verify with brain status / brain list (titles and counts), not brain show on each new document — that would be a third pass over the same text for no new information.
• Don't summarize the transcript yourself on the way through. The post-ingest hook already summarizes on local Ollama, off your context. Pipe the raw body and let it run.
• If the user only wants to know what was discussed — not to store it — ingest is the wrong tool. Ingest once, then query with consult-brain (brain search --brief, brain recall --budget N), where the read is bounded.

Note that no ingest command takes --json; they report counts, not payloads, so there is nothing here to parse and nothing to trim. (The verification commands below — brain list, brain show — do take it.) The cost is entirely in the content you carry to reach ingest-stdin.

Tags on ingest

Every ingest command takes one or more --tag flags. Tags are auto-normalized at write time (casefold-lowercase, hyphen-separated, deduped) — any case/separator at the CLI is fine.

brain ingest <file> --tag interview-prep --tag career-2026

Disabling the post-ingest summary

All four ingest commands run an LLM auto-summary post-hook by default (Q1-D). Skip with --no-enrich when:

• The doc is tiny and the summary would be noisy.
• Ollama is known down and you want the ingest to complete fast.
• You're scripting a bulk import and will run brain enrich --backfill afterwards.

brain ingest-stdin --source slack --external-id ... --title ... --no-enrich

Verifying after ingest

After any non-trivial ingest, sanity-check:

brain status                                   # totals + breakdown by source
brain list --source krisp --limit 5            # see the freshest rows

Those two are titles and counts — cheap, and enough to confirm the rows landed. Only if something looks wrong, spot-check one document with brain show <id-prefix>. Do not brain show each thing you just ingested: you already had that text in context on the way in, and re-reading a batch of transcripts costs more than the ingest did.

If brain doctor flags an Ollama / Postgres issue, route to brain-maintenance.

Safety rules

• Never bulk-ingest the inbox. brain ingest-gmail enforces this with the scope-flag requirement; respect the spirit when picking scopes.
• Jira: the user's own issues, time-bounded, confidential. No JQL without a time bound, no project-wide pulls, always --sensitivity confidential.
• GitHub: the user's own activity, date-bounded. Never an organisation-wide pull; a repository that is not public is always confidential.
• Don't invent content for stdin ingests. If the MCP returned nothing, tell the user — don't generate a placeholder transcript.
• No untagged dumps. Always pass --tag so the user can find/clean the batch later.
• Mind external_id stability. A Krisp meeting that gets re-ingested with a different --external-id becomes a duplicate row. Match what the MCP returns as the canonical meeting id.